KR101634979B1 - Model based prediction in a critically sampled filterbank - Google Patents

Abstract

The present disclosure relates to audio source coding systems. In particular, the present disclosure relates to audio source coding systems that use linear prediction in combination with a filter bank. A method for estimating a first sample 615 of a first subband signal of a first subband of an audio signal is described. The first subband signal of the audio signal is determined using an analysis filterbank 612 that includes a plurality of analysis filters each providing a plurality of subband signals of a plurality of subbands from the audio signal. The method includes determining a model parameter (613) of a signal model; Based on the signal model, a prediction coefficient to be applied to the previous sample 614 of the first decoded subband signals derived from the first subband signal, based on the model parameter 613 and based on the analysis filter bank 612, Wherein the time slot of the previous sample (614) is before the type slot of the first sample (615); And determining an estimate of the first sample 615 by applying a prediction coefficient to the previous sample 614.

Description

MODEL BASED PREDICTION IN A CRITICALLY SAMPLED FILTERBANK IN THRUALLY SAMPLED FILTER BANK.

The present invention relates to audio source coding systems. More particularly, the present invention relates to audio source coding systems that use linear prediction in combination with a filter bank.

There are two important signal processing tools applied to the system for source coding of audio signals, namely, thresholded sampled filter banks and linear prediction. Critically sampled filter banks (e.g., modulated discrete cosine transform (MDCT) -based filter banks) allow direct access to time-frequency representations where perceptual indifference and signal redundancy can be exploited. Linear prediction enables efficient source modeling of audio signals, particularly speech signals. The combination of the two tools, i. E. The use of prediction in the subbands of the filter bank, was mainly used for high bit rate audio coding. For low bit rate coding, the challenge by prediction in subbands is to keep the cost (i.e. bit rate) of the predictors' descriptions low. Another challenge is to control the resulting noise shaping of the prediction error signal obtained by the subband predictor.
US2006 / 0015329A1 describes a method for audio coding using a waveform synthesizer to generate a set of predicted samples of an audio signal.

For the challenge of encoding the description of the subband predictor in a bit-efficient manner, the possible path is to estimate the predictor from the previously decoded portions of the audio signal and thereby avoid the cost of the predictor description altogether. If the predictor can be determined from previously decoded portions of the audio signal, the predictor can be determined at the encoder and at the decoder without having to send the predictor description from the encoder to the decoder. This approach is called the backward adaptive prediction scheme. However, the backward-adaptive prediction scheme generally degrades significantly when the bit rate of the encoded audio signal is reduced. An alternative or additional path to efficient encoding of the subband predictor is to identify a more natural predictor description, e.g., a description using a unique structure of the audio signal to be encoded. For example, low bitrate rate coding generally applies a forward adaptive scheme based on a simple representation of a short-term predictor (using short-term correlations) and a long-term predictor (using long-term corrections by the fundamental pitch of the rate signal) do.

For the challenge of controlling the noise shaping of the predictive error signal, while the noise shaping of the predictor can be well controlled within the subband, the final output audio signal of the encoder is transformed into aliased artifacts (substantially smooth spectral noise shaping (Except for the audio signals that represent it).

An important case of the subband predictor is the implementation of long-term prediction in a filter bank with overlapping windows. The long-term predictor can be described as a single or a small number of predictive parameters, typically using redundancies in periodic and nearly periodic audio signals (e.g., velocity signals representing inherent pitch). The long-term predictor may be defined at successive times by a delay reflecting the periodicity of the audio signal. When this delay is large compared to the length of the filter bank window, the long-term predictor can be executed in a separate time domain by shift or partial delay and can be transformed back into a causal predictor in the subband domain. These long-term predictors do not generally represent aliasing artifacts, but there are significant disadvantages in the computational complexity caused by the need for additional filter bank operations for conversion from the time domain to the subband domain. Also, the manner of determining the delay in the time domain and converting the delay to the subband predictor is not applicable for cases where the period of the audio signal to be encoded is similar to or less than the filter bank window size.

The present invention describes the above-mentioned disadvantages of subband prediction. In particular, the present invention describes methods and systems that allow for bit-rate efficient description of subband predictors and / or allow reduction of aliasing artifacts caused by subband predictors. In particular, the methods and systems described herein enable the implementation of low bit rate audio coders that use subband prediction to cause reduced levels of aliasing artifacts.

The present invention describes methods and systems for improving the quality of audio source coding employing prediction in the subband domain of a critically sampled filter bank. The methods and systems may use a simple description of the subband predictors, and the description is based on signal models. Alternatively or additionally, the methods and systems may utilize an efficient implementation of direct predictors in the subband domain. Alternatively or additionally, the methods and systems may utilize the cross-subband predictor terms described herein to allow reduction of aliasing artifacts.

As outlined in the present invention, a simple description of the subband predictors is based on the frequency of the sine wave, the period of the periodic signal, the slightly non-harmonic spectrum experienced by the vibrations of the tight string, and / . ≪ / RTI > For the case of the long-term predictor, the cyclic signal model is shown to provide high quality causal predictors for a range of lag parameters (or delays) that contain values that are shorter and / or longer than the window size of the filter bank . This means that the periodic signal model can be used to implement the long-term subband predictor in an efficient manner. Seamless motion is provided as an approximation of arbitrary delay from sine wave model based prediction. Direct implementations of predictors in the subband domain enable explicit access to perceptual features of the generated quantization distortions. In addition, the implementation of predictors in the subband domain enables access to numerical properties such as the prediction gain for the parameters and the dependence of the predictors. For example, a signal model based analysis may be useful in that the prediction gain is important only in a subset of the considered subbands, and that variations in the predictor coefficients as a function of the parameters selected for transmission may be helpful in the design of the parameter formats, . Moreover, it can be shown that the computational complexity can be significantly reduced compared to predictor implementations that rely on the use of algorithms that operate both in the time domain and in the subband domain. In particular, the methods and systems described herein can be used to perform subband prediction directly in the subband domain without having to determine and apply predictors (e.g., long term delay) in the time domain.

The use of cross-subband periods in the subband predictors enables significantly improved frequency domain noise shaping properties as compared to in-band predictors (which depend only on in-band prediction). By doing this aliasing artifacts can be reduced, thereby enabling the use of subband prediction for relatively low bit rate audio coding systems.

According to an aspect, a method for estimating a first sample of a first subband of an audio signal is described. The first subband of the audio signal could be determined from the audio signal using an analysis filter bank comprising a plurality of analysis filters each providing a plurality of subband signals in a plurality of subbands. The time domain audio signal may be submitted to an analysis filter bank, thereby producing a plurality of subband signals in a plurality of subbands. Each of the plurality of subbands typically includes an audio signal in a different frequency range, thereby providing access to different frequency components of the audio signal. The plurality of subbands may have equal or equal subband spacing. The first subband corresponds to one of the plurality of subbands provided by the analysis filter bank.

The analysis filter bank may have various properties. A synthesis filter bank comprising a plurality of synthesis filters may have similar or identical properties. The attributes described for the analysis filterbank and analysis filters are also applicable to the properties of the synthesis filterbank and synthesis filters. In general, the combination of the analysis filter bank and the synthesis filter bank allows a complete reconstruction of the audio signal. The analysis filters of the analysis filter bank may be shift-invariant with respect to each other. Alternatively or additionally, the analysis filters of the analysis filter bank may include a common window function. In particular, the analysis filters of the analysis filter bank may contain different modulated versions of the common window function. In one embodiment, the common window function is modulated using a cosine function, thereby yielding a cosine modulated analysis filter bank. In particular, the analysis filter bank may (or may correspond to) one or more of MDCT, QMF, and / or ELT transformations. The common window function may have a finite duration (K). The duration of the common window function may be that successive samples of the subband signal are determined using overlapping segments of the time domain audio signal. As such, the analysis filter bank may include nested transforms. The analysis filters of the analysis filter bank may form an orthogonal and / or orthogonal basis. As another attribute, the analysis filter bank may correspond to a critically sampled filter bank. In particular, the number of samples of the plurality of subband signals may correspond to the number of samples of the time domain audio signal.

The method may include determining a model parameter of the signal model. It should be noted that the signal model can be described using a plurality of model parameters. As such, the method may include determining a plurality of model parameters of the signal model. The model parameter (s) may be extracted from the received bitstream that includes or represents the model parameters and the prediction error signal. Alternatively, the model parameter (s) may be determined (e.g., based on frame-by-frame) by fitting the signal model to the audio signal, e.g., using a mean square error scheme.

The signal model may include one or more sine wave model components. In this case, the model parameter may represent one or more frequencies of one or more sine wave model components. By way of example, the model parameters may represent the fundamental frequency (?) Of a multi-sinusoidal signal model and the multi-sinusoidal signal includes sinusoidal model components at frequencies corresponding to multiples of the fundamental frequency (?) . As such, the multi-sinusoidal signal model may comprise a periodic signal component, wherein the periodic signal component comprises a plurality of sinusoidal components and the plurality of sinusoidal components have a frequency that is a multiple of the fundamental frequency ([Omega]). As shown herein, this periodic signal component can be used to model the delay in the time domain (e.g., used for long-term predictors). The signal model may include one or more model parameters indicative of a shift and / or a deviation of the signal model from the periodic signal model. The shift and / or deviation may represent a deviation of frequencies of a plurality of sinusoidal components of the periodic signal model from respective multiple (q?) Of the fundamental frequency (?).

The signal model may comprise a plurality of periodic signal components. Each of the periodic signal components may be described using one or more model parameters. The model parameters may represent a plurality of fundamental frequencies (Ω ₀ , Ω ₁ , ..., Ω _M-1 ) of a plurality of periodic signal components. Alternatively or additionally, the signal model may be described by a predetermined and / or adjustable mitigation parameter (which may be one of the model parameters). The relaxation parameter may be configured to stabilize or planarize the linear spectrum of the periodic signal component. Specific examples of signal models and associated model parameters are described in the Examples section herein.

The model parameter (s) may be determined such that the mean value of the squared prediction error signal is reduced (e. G., Minimized). The prediction error signal may be determined based on the difference between the estimate of the first sample and the first sample. In particular, the mean value of the squared prediction error signal may be determined based on a plurality of successive first samples of the first subband signal and based on a corresponding plurality of estimated first samples. In particular, it is proposed herein to model at least a first subband signal of an audio signal or an audio signal using a signal model described by one or more model parameters. The model parameters are used to determine one or more prediction coefficients of a linear predictor that determines a first estimated subband signal. The difference between the first subband signal and the first estimated subband signal yields a predicted error subband signal. The one or more model parameters may be determined such that the mean value of the squared predicted error subband signal is reduced (e.g., minimized).

The method may further comprise determining a prediction coefficient to be applied to a previous sample of the first decoded subband signal derived from the first subband signal. In particular, the previous sample may be determined by adding a (quantized version) of the prediction error signal to the corresponding sample of the first subband signal. The first decoded subband signal may be the same as the first subband signal (e.g., for a lossless encoder). The time slot of the previous sample is typically before the time slot of the first sample. In particular, the method may comprise determining one or more prediction coefficients of a repeated (finite impulse response) prediction filter configured to determine a first sample of a first subband signal from one or more previous samples.

The one or more prediction coefficients may be determined based on the signal model, based on the model parameters, and on the analysis filter bank. In particular, the prediction coefficients may be determined based on the analytical evaluation of the signal model and the analysis filter bank. Analysis of the signal model and analysis The analysis of the filterbank may result in the determination of the reference table and / or the analysis function. As such, the prediction coefficients may be determined using reference tables and / or analytic functions, and the reference tables and / or analytic functions may be predetermined based on the signal model and based on the analysis filter banks. The lookup table and / or analysis function may provide the prediction coefficient (s) as a function of the parameters derived from the model parameter (s). The parameters derived from the model parameters can be, for example, model parameters or can be obtained from the model parameters using a predetermined function. As such, the one or more prediction coefficients may be computed in an efficient manner using an analytical function that provides one or more prediction coefficients (only) in dependence of (more than) one or more parameters derived from one or more model parameters and / . ≪ / RTI > Thus, the determination of the prediction coefficients can be reduced by a simple search of the entries in the look-up table. As indicated above, the analysis filter bank may include or represent a modulated structure. As a result of this modulated structure, it is observed that the absolute value of one or more prediction coefficients is independent of the index number of the first subband. This means that the look-up table and / or analytic function may be shift-invariant (except signal values) with respect to the index number of the plurality of subbands. In these cases, the parameters derived from the model parameters, i.e., the parameters input to the look-up table and / or the analysis function to determine the prediction coefficients, may include model parameters in a relative manner with respect to one of the plurality of subbands . ≪ / RTI >

)의 함수로서) 제공하도록 미리 결정될 수 있다. 이와 같이, 예측 계수는 결정된 상대적인 오프셋을 사용하여 참조표에 기초하여 및/또는 분석 함수에 기초하여 결정될 수 있다. 미리 결정된 참조표는 제한된 수의 가능한 상대적인 오프셋들에 대해 제한된 수의 엔트리들을 포함할 수 있다. 이러한 경우에서, 결정된 상대적인 오프셋은 참조표로부터 예측 계수를 검색하기 전에 제한된 수의 가능한 상대적인 오프셋들로부터 가장 가까운 가능한 상대적인 오프셋으로 근사될 수 있다.As described above, the model parameters may represent the fundamental frequency (Ω) of a multi-sinusoidal signal model (eg, of a periodic signal model). In these cases, the step of determining the prediction coefficients may comprise determining a multiple of the fundamental frequency (OMEGA) lying within the first subband. If a multiple of the fundamental frequency (?) Is within the first sub-band, the relative offset of a multiple of the fundamental frequency (?) From the center frequency of the first sub-band may be determined. In particular, the relative offset of a multiple of the fundamental frequency (?) Closest to the center frequency of the first sub-band can be determined. The reference table and / or analysis function may be used to determine whether the look-up table and / or the analysis function has a predictive coefficient as a function of the relative offsets possible from the center frequency of the subband (e.g., the normalized frequency f As a function and / or a shift parameter (

) ≪ / RTI > As such, the prediction coefficients may be determined based on the lookup table using the determined relative offset and / or based on the analysis function. The predetermined reference table may include a limited number of entries for a limited number of possible relative offsets. In such a case, the determined relative offset can be approximated to a closest possible relative offset from a limited number of possible relative offsets before retrieving the prediction coefficients from the look-up table.

On the other hand, if no multiple of the fundamental frequency? Is within the first sub-band, i.e. rather within the extended frequency range surrounding the first sub-band, then the prediction coefficient is set to zero. In these cases, the estimate of the first sample may also be zero.

))에 의존하는 것이 본 발명에서 보여진다. 이와 같이, 복수의 참조표들이 주기성(T)의 복수의 상이한 값들에 대해 미리 결정될 수 있다. 모델 파라미터(즉, 주기성(T))는 복수의 참조표들 중 적절한 것을 선택하기 위해 사용될 수 있고, 예측 계수는 (상대적인 오프셋을 사용하여, 예를 들면, 시프트 파라미터(

)를 사용하여) 복수의 참조표들 중 선택된 것에 기초하여 결정될 수 있다. 이와 같이, 상대적으로 높은 정확도를 가질 수 있는 모델 파라미터(예를 들면, 주기성(T)을 나타내는)는 감소된 정확도의 한 쌍의 파라미터들(예를 들면, 주기성(T) 및 상대적인 오프셋)로 디코딩될 수 있다. 한 쌍의 파라미터들의 제 1 파라미터(예를 들면, 주기성(T))는 특정한 참조표를 선택하기 위해 사용될 수 있고, 제 2 파라미터(예를 들면, 상대적인 오프셋)는 선택된 참조표내 엔트리를 식별하기 위해 사용될 수 있다.The step of determining the prediction coefficients may comprise selecting one of the plurality of look-up tables based on the model parameters. By way of example, the model parameter may represent the fundamental frequency (?) Of the periodic signal model. The fundamental frequency (?) Of the periodic signal model corresponds to the periodicity (T) of the periodic signal model. In the case of relatively small periodicity (T), it is shown herein that the periodic signal model converges towards a single sinusoidal model. Also, in the case of relatively large periodicity (T), the look-up tables vary slowly according to the absolute value of T and are mainly used for relative offset (i.e., shift parameter

)) Is shown in the present invention. As such, a plurality of look-up tables may be predetermined for a plurality of different values of the periodicity T. The model parameters (i.e., periodicity T) may be used to select the appropriate one of the plurality of look-up tables, and the prediction coefficients may be calculated using the relative offset (e.g.,

)) May be determined based on the selected one of the plurality of reference tables. As such, a model parameter (e.g., representing periodicity T) that can have a relatively high accuracy can be decoded (e.g., periodicity T and relative offset) with a pair of reduced accuracy parameters . The first parameter (e.g., periodicity T) of the pair of parameters may be used to select a particular reference table and the second parameter (e.g., relative offset) may be used to identify the selected reference in- Can be used.

The method may further comprise determining an estimate of the first sample by applying a prediction coefficient to the previous sample. Applying the prediction coefficient to the previous sample may include multiplying the prediction coefficient by the value of the previous sample, thereby calculating an estimate of the first sample. In general, a plurality of first samples of a first subband signal are determined by applying a prediction coefficient to a series of previous samples. The step of determining an estimate of the first sample may further comprise applying a scaling gain to the prediction coefficient and / or the first sample. The scaling gain or its indication may be used, for example, for long term prediction (LTP). In other words, the scaling gain may result from a different predictor (e.g., from a long-term predictor). The scaling gain may be different for different subbands. In addition, the scaling gain may be transmitted as part of the encoded audio signal.

As such, an efficient description of the subband predictor (including one or more prediction coefficients) is provided by using the signal model described by the model parameters. The model parameters are used to determine one or more prediction coefficients of the subband predictor. This means that the audio encoder needs to transmit an indication of the model parameters, rather than an indication of one or more prediction coefficients. In general, model parameters can be encoded more efficiently (i.e., with fewer bits) than one or more prediction coefficients. Thus, the use of model-based prediction enables low bit rate subband encoding.

The method may further comprise determining a prediction mask representing a plurality of previous samples in the plurality of prediction mask support subbands. The plurality of prediction mask support subbands may include at least one of a plurality of subbands different from the first subbands. As such, the subband predictor can be configured to estimate a sample of the first subband signal from samples of one or more other subband signals from a plurality of subband signals different from the first subband signal. This is referred to herein as cross-subband prediction. The prediction mask includes a plurality of previous samples used to estimate a first sample of the first subband signal (e.g., a time lag with respect to a time slot of the first sample and / Subband index lag). ≪ / RTI >

The method may proceed to determining a plurality of prediction coefficients to be applied to a plurality of previous samples. The plurality of prediction coefficients may be determined based on the model parameters, based on the signal model, and on an analysis filter bank (e.g., using model-based prediction schemes as outlined above and herein) . As such, the plurality of prediction coefficients may be determined using one or more model parameters. In other words, a limited number of model parameters may be sufficient to determine a plurality of prediction coefficients. This means that by using model-based subband prediction, cross-subband prediction can be performed in a bit-rate efficient manner.

The method may include determining an estimate of the first sample by applying each of the plurality of prediction coefficients to a plurality of previous samples. The step of determining an estimate of the first sample generally includes determining a sum of a plurality of previous samples weighted by the plurality of respective prediction coefficients.

As described above, the model parameter may represent the periodicity (T). The plurality of look-up tables used to determine one or more prediction coefficients may include look-up tables for different values of the periodicity (T). In particular, the plurality of look-up tables may include look-up tables for different values of the periodicity (T) of the predetermined step size? T within the range of [T _min , T _max ]. As outlined herein, T _min can be in the range of 0.25, and T _max can be in the range of 2.5. T _min can be selected such that for T < T _min , the audio signal can be modeled using a signal model that includes one sinusoidal model component. T _max is, T> with respect to the T _max, the periodicity references to (T _max to T _max + 1) Table that the periodicity - be selected to substantially correspond to the reference schedule for the (T _max 1 to T _max) . In general, the same applies to the periodicity ( _Tmax + n to _Tmax + n + 1 (usually for n? 0)).

)에 대응할 수 있거나 또는 그로부터 도출될 수 있다.The method may include determining the selected reference table as a look-up table for the periodicity (T) as a model parameter. After having a selected reference table that includes or displays one or more prediction coefficients, the reference parameter may be used to identify one or more appropriate entries in the selected reference table that each represent one or more prediction coefficients. The reference parameter is a shift parameter (

) Or may be derived therefrom.

Method, the periodicity (T> T _max) for with respect to representing the model parameters, and the remaining periodicity (T _r) is a range-rest periodicity by subtracting the integer value from T to lie in [T _max 1, T _max] (T _r) . The look-up table for determining the prediction coefficients may then be determined as a look-up table for the remaining periodicity (T _r ).

))는 비율(T_min/T)에 따라 크기 조정될 수 있다. 하나 이상의 예측 계수들은 이후 선택된 참조표 및 크기 조정된 참조 파라미터를 사용하여 결정될 수 있다. 특히, 하나 이상의 예측 계수들은 크기 조정된 참조 파라미터에 대응하는 선택된 참조표의 하나 이상의 엔트리들에 기초하여 결정될 수 있다.The method may include selecting a look-up table to determine, for a model parameter representing periodicity (T < T _min ), one or more prediction coefficients as a look-up table for periodicity (T _min ). Further, a reference parameter (e.g., a shift parameter (e.g., a reference value) for identifying one or more entries of a selected reference table that provides one or more prediction coefficients

) Can be scaled according to the ratio _Tmin / T. The one or more prediction coefficients may then be determined using the selected reference table and the scaled reference parameter. In particular, the one or more prediction coefficients may be determined based on one or more entries of the selected reference table corresponding to the scaled reference parameter.

As such, the number of look-up tables can be limited to a predetermined range [ _Tmin , _Tmax ], thereby defining the memory requirements of the audio encoder / decoder. Nevertheless, the prediction coefficients can be determined for all possible values of the periodicity T using predetermined reference tables, thereby enabling a computationally efficient implementation of the audio encoder / decoder.

According to another aspect, a method for estimating a first sample of a first subband signal of an audio signal is described. As outlined above, the first subband signal of the audio signal is determined using an analysis filter bank comprising a plurality of analysis filters each providing a plurality of subband signals in a plurality of subbands from the audio signal . The features described above are also applicable to the method described below.

The method includes determining a prediction mask representing a plurality of previous samples in a plurality of prediction mask supported subbands. The plurality of prediction mask support subbands include at least one of a plurality of subbands different from the first subbands. In particular, the plurality of prediction mask capable subbands may comprise a first subband and / or the plurality of prediction mask capable subbands may comprise one or more of a plurality of subbands directly adjacent to the first subband .

The method may further comprise determining a plurality of prediction coefficients to be applied to a plurality of previous samples. A plurality of previous samples may generally be derived from a plurality of subband signals of the audio signal. In particular, a plurality of previous samples generally correspond to samples of a plurality of decoded subband signals. The plurality of prediction coefficients may correspond to prediction coefficients of a repeated (finite impulse response) prediction filter that also considers one or more samples of subbands different from the first subband. The estimate of the first sample may be determined by applying a plurality of prediction coefficients to a plurality of previous samples, respectively. As such, the method enables subband prediction using one or more samples from different (e.g., adjacent) subbands. By doing this, the aliasing artifacts caused by the subband prediction based coders can be reduced.

The method may further comprise determining a model parameter of the signal model. The plurality of prediction coefficients may be determined based on the model parameters and based on the analysis filter bank, based on the signal model. As such, the plurality of prediction coefficients may be determined using the model-based prediction described herein. In particular, the plurality of prediction coefficients may be determined using reference tables and / or analytic functions. The reference table and / or analysis function may be predetermined based on the signal model and based on the analysis filter bank. In addition, the look-up table and / or analysis function may provide a plurality of prediction coefficients as a function of (only) parameters derived from the model parameters. Thus, the model parameters can directly provide a plurality of prediction coefficients using a look-up table and / or analytic function. As such, the model parameters can be used to efficiently describe the coefficients of the cross-subband predictor.

According to another aspect, a method for encoding an audio signal is described. The method may comprise determining a plurality of subband signals from the audio signal using an analysis filter bank comprising a plurality of analysis filters. The method may proceed to estimating samples of a plurality of subband signals using any one of the prediction methods described herein, thereby yielding a plurality of estimated subband signals. In addition, the plurality of samples of the prediction error subband signals may be determined based on the corresponding samples of the plurality of subband signals and samples of the plurality of estimated subband signals. The method may proceed to quantizing a plurality of prediction error subband signals and generating an encoded audio signal. The encoded audio signal may (e.g., contain) a plurality of quantized prediction error subband signals. In addition, the encoded signal may represent (e.g., include) one or more parameters used to estimate samples of a plurality of estimated subband signals and may include, for example, a plurality of estimated subband signals Lt; / RTI &gt; may be used to determine one or more predictor coefficients that are subsequently used to estimate the samples of the model.

According to another aspect, a method for decoding an encoded audio signal is described. The encoded audio signal may generally represent one or more parameters to be used to estimate a plurality of quantized prediction error subband signals and samples of a plurality of estimated subband signals. The method may include semi-quantizing a plurality of quantized prediction error subband signals, thereby yielding a plurality of semi-quantized prediction error subband signals.

The method may also include estimating samples of the plurality of estimated subband signals using any one of the prediction methods described herein. Samples of the plurality of decoded subband signals may be determined based on corresponding samples of the plurality of estimated subband signals and based on samples of the plurality of semi-quantized prediction error subband signals. The decoded audio signal may be determined from a plurality of decoded subband signals using a synthesis filter bank comprising a plurality of synthesis filters.

According to another aspect, a system configured to estimate one or more first samples of a first subband signal of an audio signal is described. The first subband signal of the audio signal may be determined using an analysis filter bank comprising a plurality of analysis filters providing a plurality of subband signals from the audio signal in a plurality of respective subbands. The system may include a predictor calculator configured to determine model parameters of the signal model. The predictor calculator may also be configured to determine one or more prediction coefficients to be applied to one or more previous samples of the first decoded subband signal derived from the first subband signal. As such, the predictor calculator can be configured to determine one or more prediction coefficients of the repeated prediction filter, especially of the repeated subband prediction filter. The one or more prediction coefficients may be determined based on the model parameters based on the signal model and on an analysis filter bank (e.g., using model-based prediction methods described herein). The timeslots of one or more previous samples generally precede the timeslots of one or more first samples. The system may further comprise a subband predictor configured to determine an estimate of one or more first samples by applying one or more prediction coefficients to one or more previous samples.

According to another aspect, a system configured to estimate one or more first samples of a first subband signal of an audio signal is described. The first subband signal corresponds to a first one of the plurality of subbands. The first subband signal is typically determined using an analysis filter bank that includes a plurality of analysis filters each of which provides a plurality of subband signals for a plurality of subbands. The system includes a predictor calculator configured to determine a prediction mask representing a plurality of previous samples in a plurality of prediction mask supported subbands. The plurality of prediction mask support subbands include at least one of a plurality of subbands different from the first subbands. The predictor calculator is also configured to determine a plurality of prediction coefficients (or repeated prediction filters) to be applied to a plurality of previous samples. In addition, the system includes a subband predictor configured to determine an estimate of one or more first samples by applying a plurality of prediction coefficients to a plurality of previous samples, respectively.

According to another aspect, an audio encoder configured to encode an audio signal is described. The audio encoder includes an analysis filter bank configured to determine a plurality of subband signals from the audio signal using a plurality of analysis filters. The audio encoder also includes a predictor calculator and a subband predictor described herein configured to estimate samples of a plurality of subband signals, thereby producing a plurality of estimated subband signals. The encoder may also include a difference unit configured to determine samples of the plurality of prediction error subband signals based on the plurality of subband signals and the corresponding samples of the plurality of estimated subband signals. The quantization unit may be used to quantize a plurality of prediction error subband signals. In addition, the bitstream generation unit may be configured to generate a plurality of quantized prediction error subband signals and one or more parameters (e.g., one or more model parameters) used to estimate samples of a plurality of estimated subband signals And generate an encoded audio signal.

According to another aspect, an audio decoder configured to decode an encoded audio signal is described. The encoded audio signal represents (e.g., includes) a plurality of quantized prediction error subband signals and one or more parameters used to estimate samples of a plurality of estimated subband signals. The audio decoder may comprise an inverse quantizer configured to semi-quantize a plurality of quantized prediction error subband signals, thereby producing a plurality of semi-quantized prediction error subband signals. The decoder also includes a predictor calculator and a subband predictor described herein configured to estimate samples of a plurality of estimated subband signals. The summation unit may be used to determine samples of the plurality of decoded subband signals based on the corresponding samples of the plurality of estimated subband signals and based on the samples of the plurality of semi-quantized prediction error subband signals . The synthesis filter bank may also be used to determine the decoded audio signal from a plurality of decoded subband signals using a plurality of synthesis filters.

According to another aspect, a software program is described. A software program may be adapted to perform the method steps outlined herein for execution on a processor and when executed on a processor.

According to another aspect, a storage medium is described. The storage medium may comprise a software program adapted for performing the method steps described herein in outline for execution on a processor and on a processor.

According to another aspect, a computer program product is described. A computer program, when executed on a computer, may include executable instructions for performing the method steps outlined herein.

It should be noted that the methods and systems comprising their preferred embodiments outlined in this patent application may be used alone or in combination with other methods and systems described herein. In addition, all aspects of the methods and systems outlined in this patent application may be combined arbitrarily. In particular, the features of the claims may be combined with one another in any manner.

The present invention describes methods and systems for improving the quality of audio source coding employing prediction in the subband domain of a critically sampled filter bank. The methods and systems may utilize an efficient implementation of direct predictors in the subband domain. Alternatively or additionally, the methods and systems may utilize the cross-subband predictor terms described herein to allow reduction of aliasing artifacts.

1 is a block diagram of an exemplary audio decoder for applying linear prediction in a filter bank domain (i.e., in a subband domain);
Figure 2 illustrates exemplary prediction masks in a time frequency grid.
3 shows an exemplary statistical table for a sine wave model based predictor calculator.
4 is a diagram illustrating exemplary noise shaping resulting from in-band subband prediction;
5 illustrates an exemplary noise shaping resulting from a cross-band subband prediction;
FIG. 6A illustrates an exemplary two-dimensional quantization grid based on a statistical table for periodic model-based predictor computation. FIG.
Figure 6B illustrates the use of different prediction masks for different ranges of signal periodicity;
Figures 7A and 7B are flowcharts of exemplary encoding and decoding methods using model-based subband prediction.

The present invention is described below as illustrative examples without limiting the scope or spirit of the invention with reference to the accompanying drawings.

The embodiments described below are merely illustrative of the principles of the present invention for model-based prediction in a critically sampled filter bank. Variations and modifications of the devices and details described herein will be apparent to those skilled in the art. It is, therefore, intended to be limited only by the scope of the pending patent claims, and not limited by the specific details expressed as the description and the description of the embodiments herein.

Figure 1 shows a block diagram of an exemplary audio decoder 100 that applies linear prediction in a filterbank domain (also called a subband domain). The audio decoder 100 receives information about the prediction error signal (also referred to as the residual signal) and possibly about the description of the predictor used by the corresponding encoder to determine the prediction error signal from the original input audio signal And receives the bitstream including the bitstream. The information about the prediction error signal may be for subbands of the input audio signal and the information about the description of the predictor may be for one or more server band predictors.

Considering the received bitstream information, the dequantizer 101 may output samples 111 of the predicted error subband signals. These samples may be added to the output 112 of the subband predictor 103 and the sum 113 may be added to the output of the subband predictor 103 to maintain the recording of the previously decoded samples 113 of the subband of the decoded audio signal. (104). The output of subband predictor 103 may be referred to as estimated subband signals 112. The decoded samples 113 of the subbands of the decoded audio signal may be submitted to a synthesis filter bank 102 that transforms the subband samples into the time domain so that the time domain samples 114 of the decoded audio signal ).

In other words, the decoder 100 may operate in the subband domain. In particular, the decoder 100 may use the subband predictor 103 to determine a plurality of estimated subband signals 112. In addition, the decoder 100 may use a dequantizer 101 to determine a plurality of remaining subband signals 111. Each of a plurality of estimated subband signals 112 and a plurality of pairs of remaining subband signals 111 may be added to produce a corresponding plurality of decoded subband signals 113. A plurality of decoded subband signals 113 may be submitted to the synthesis filter bank 102 to produce a time domain decoded audio signal 114.

In one embodiment of the subband predictor 103, a given sample of a given estimated subband signal 112 is at a different time and at a different frequency than a given sample of a given estimated subband signal 112 (i.e., ) &Lt; / RTI &gt; of the sub-band samples of the buffer 104 corresponding to the sub-band samples. In other words, at a first time instant and in a first subband, a sample of the estimated subband signal 112 is associated with a second time instant (different from the first instant in time) and a second subband May be determined based on one or more samples of the decoded subband signals 113 associated with the received signal. The prediction coefficients and a collection of their adjuncts to the time and frequency masks may define the predictor 103 and this information may be supplied by the predictor calculator 105 of the decoder 100. [ The predictor calculator 105 outputs information defining the predictor 103 by conversion of the signal model data included in the received bitstream. An additional gain may be transmitted that alters the scaling of the output of the predictor 103. [ In one embodiment of the predictor calculator 105, the signal model data is provided in the form of an efficient parameterized line spectrum, and each line of the parameterized line spectrum, or a subsequent group of parameterized line spectra The lines are used to indicate the statistical values of the predictor coefficients. As such, the signal model data provided in the received bitstream may be used to identify predetermined reference entries, and entries from the look-up table may include predictor coefficients (also referred to as prediction coefficients) to be used by the predictor 103 &Lt; / RTI &gt; The methodology applied in the reference table may depend on the balance between complexity and memory requirements. For example, the nearest neighborhood search may be used to achieve the lowest complexity, while the interpolation search method may provide similar performance with a smaller table size.

As indicated above, the received bitstream may include one or more explicitly transmitted gains (or indications of explicitly transmitted gains). The gains may be applied as part of the predictor operation or thereafter. The one or more explicitly transmitted gains may be different for different subbands. The (explicitly transmitted) additional gains (indicia of) are provided in addition to one or more model parameters used for the predictor coefficients determined by the predictor 103. As such, the additional gains can be used to scale the prediction coefficients of the predictor 103.

Figure 2 shows exemplary prediction mask supports in a time frequency grid. Prediction mask supports may be used for predictors 103 operating in a filter bank with an equal time frequency resolution, such as a cosine modulated filter bank (e.g., an MDCT filter bank). The notation is shown in diagram 201, i.e., the target dark shaded subband sample 211 is the output of the prediction based on the lightly shaded subband sample 212. In diagrams 202-205, the set of lightly shaded subband samples represents predictor mask support. The combination of source subband samples 212 and target subband samples 211 will be referred to as prediction mask 201. [ The time-frequency grid can be used to align subband samples near the target subband samples. The time slot index is increasing from left to right and the subband frequency index is increasing from below to above. It should be noted that FIG. 2 illustrates exemplary cases of prediction masks and predictor mask supports and a number of other prediction masks and predictor mask supports may be used. Exemplary prediction masks include:

The prediction mask 202 is used to estimate the subband samples 221 in the band of estimated subband samples 221 at time instants k from the two previously decoded subband samples 222 at time instants k- Define the forecast.

The prediction mask 203 generates a time instant k based on three previous decoded subband samples 232 at time instant k-1 and in subband n-1, n, n + Band prediction of subband samples 231 estimated at subband n and at subband n.

The prediction mask 204 is used to estimate the time instant k based on the three previous decoded subband samples 242 at time instant k-1 and in subband n-1, n, n + Band prediction of three estimated subband samples 241 in three different subbands (n-1, n, n + 1). The cross-band prediction is determined based on all of the three previous decoded subband samples 242 in subbands (n-1, n, n + 1) Can be performed.

The prediction mask 205 may be applied to the twelve previously decoded subbands at time instants k-2, k-3, k-4, k-5 and in subbands n- Band prediction of subband samples 251 estimated in subband n and at time instant k based on samples 252. [

Figure 3 shows a statistical table for a sine wave model based predictor calculator 105 operating in a cosine modulation filter bank. The prediction mask support is in diagram 204. For a given frequency parameter, a subband with the closest subband center frequency may be selected as the central target subband. The difference between the frequency parameter and the center frequency of the central target subband may be calculated as the bins of the frequency spacing of the filter bank. This provides a value between -0.5 and 0.5, which can be approximated by the closest available entry in the statistical table shown in abscissa of the nine graphs 301 of FIG. Which produces a 3x3 matrix of coefficients to be applied to the most recent values of the plurality of decoded subband signals 113 in the subband buffer 104 of the target subband and its two adjacent subbands. The resulting 3x1 vector constitutes the contribution of the subband predictor 103 to these three subbands for a given frequency parameter. The process may be repeated in an additional manner for all sinusoidal components in the signal model.

In other words, FIG. 3 shows an example of a model-based description of a subband predictor. It is assumed that the input audio signal contains one or more sinusoidal components at the fundamental frequencies (? ₀ ,? ₁ , ..., _M-1 ). For each of the one or more sinusoidal components, the subband predictor may be determined using a predetermined prediction mask (e.g., prediction mask 204). The fundamental frequency (?) Of the input audio signal may be in one of the subbands of the filter bank. These subbands may be referred to as the central subbands for this particular fundamental frequency ([Omega]). The fundamental frequency (?) Can be expressed as a value in the range of -0.5 to 0.5 with respect to the center frequency of the central sub-band. The audio encoder may send information about the fundamental frequency (?) To the decoder 100. [ The predictor calculator 105 of the decoder 100 determines the coefficient 3 of the prediction matrix by determining the coefficient value 302 for each frequency value 303 of the fundamental frequency X 3 matrix can be used. This means that the coefficients for the subband predictor 103 using the prediction mask 204 can be determined using only information received for a particular fundamental frequency ([Omega]). In other words, a bit-rate efficient description of the subband predictor can be provided, for example, by modeling the input audio signal using a model of one or more sinusoidal components.

4 illustrates an exemplary noise shaping resulting from in-band subband prediction in a cosine modulation filter bank. The signal model used to perform the in-band subband prediction is a second-order auto-regressive probability process with a peak-rich resonance described by a second order differential derived by random Gaussian white noise. Curve 401 shows the magnitude spectrum measured for the realization of the process. For this example, the prediction mask 202 of FIG. 2 is applied. That is, the predictor calculator 105 provides the subband predictor 103 for a given target subband 221 based on the previous subband samples 222 only in the same subband. Substituting the inverse quantizer 101 with a Gaussian white noise generator results in a synthesized magnitude spectrum 402. As can be appreciated, strong alias artifacts arise during synthesis because the synthesis spectrum 402 contains peaks that do not match the original spectrum 401.

Figure 5 illustrates an exemplary noise shaping resulting from cross-band subband prediction. The setting is the same as the setting in Fig. 4 except that the prediction mask 203 is applied. Thus, the calculator 105 provides the predictor 103 for a given target subband 231 based on the previous subband samples 232 in the target subband and in its two adjacent subbands. As can be understood from FIG. 5, the spectrum 502 of the synthesized signal substantially coincides with the spectrum 501 of the original signal. That is, aliasing problems are substantially suppressed when using cross-band subband prediction. Thus, FIGS. 4 and 5 illustrate when using cross-band subband prediction, i.e., when predicting subband samples based on previous subband samples of one or more adjacent subbands, aliasing RTI ID = 0.0 &gt; artifacts &lt; / RTI &gt; can be reduced. As a result, subband prediction can also be applied in the context of low bitrate audio encoders without the risk of causing audible aliasing artifacts. The use of cross-band subband prediction generally increases the number of prediction coefficients. However, as shown in the environment of FIG. 3, the use of models for the input audio signal (e.g., use of a sinusoidal model or periodic model) allows efficient description of the subband predictor, Enables the use of cross-band subband prediction for coders.

Next, a description of the principles of model-based prediction in a critically sampled filter bank will be outlined with reference to FIGS. 1-6 and by adding appropriate mathematical terms.

)에 의해 결정되는 제로 평균 약하게 고정된 확률 프로세스(x(t))의 선형 예측이다. 여기서 고려될 임계적으로 샘플링된 필터뱅크들에 대한 양호한 모델로서, 이를 정규 직교 기반을 구성하는 실수값 합성 파형들(w_α(t))의 집합인

로 두자. 다시 말해서, 필터뱅크는 파형들(

)에 의해 표현될 수 있다. 시간 도메인 신호(s(t))의 서브대역 샘플들은 내적들,Possible signal model-based linear predictions are based on the assumption that their statistics are based on their autocorrelation function

0.0 &gt; x (t)) &lt; / RTI &gt; As good models for the critical sampling of the filter bank to be considered here, which is a set of real numbers in this composite waveform constituting the orthonormal base (w _α (t))

Let's do it. In other words, the filter bank includes waveforms

). &Lt; / RTI &gt; The subband samples of the time domain signal s (t)

And the signal is obtained by:

Lt; / RTI &gt;

은 확률 변수들이고, 그의 공분산 행렬(R_αβ)은 다음과 같이 자동상관 함수(r(τ))에 의해 결정되고The subband samples of process x (t)

Is a random variable, and its covariance matrix R _{?? Is} determined by the autocorrelation function r (?) As follows

Where W _αβ (τ) is the cross-correlation of the two composite waveforms.

또는 디코딩된 서브대역 샘플들

의 선형 예측은 다음으로 규정될 수 있다:Subband samples from the set

Or decoded subband samples

Can be defined as: &lt; RTI ID = 0.0 &gt;

In equation (5), set B specifies the source subband samples, i.e., set B specifies the prediction mask support. The mean value of the squared prediction error is:

And the minimum mean square error (MSE) solution is obtained by solving the standard equations for the prediction coefficients c &_lt; RTI ID = 0.0 &_gt;

로 감소된다. 표준 방정식들(7)은, 예를 들면, 레빈슨-더빈 알고리즘을 사용하여 효율적인 방식으로 풀 수 있다.When the prediction coefficients satisfy equation (7), the right term of equation (6)

. Standard equations (7) can be solved in an efficient manner, for example using the Levinson-Durbin algorithm.

)이 예측자 계산기(105)에서 도출될 수 있는 신호 모델의 파라미터 표현을 송신하는 것이 본 명세서에서 제안된다. 예를 들면, 신호 모델은 신호 모델의 자기 상관 함수(r(τ))의 파라미터 표현을 제공할 수 있다. 디코더(100)는 수신된 파라미터 표현을 사용하여 자기 상관 함수(r(τ))를 도출할 수 있고, 표준 방정식들(7)에 요구된 공분산 행렬 엔트리들을 도출하기 위해 자기 상관 함수(r(τ))를 합성 파형 교차 상관(W_αβ(τ))과 조합할 수 있다. 이들 식들은 이후 예측 계수들을 얻기 위해 풀이될 수 있다.Prediction coefficients (

) To transmit a parameter representation of the signal model that can be derived from this predictor calculator 105 is proposed herein. For example, the signal model may provide a parameter representation of the autocorrelation function of the signal model (r (tau)). The decoder 100 may derive an autocorrelation function r () using the received parameter representation and may derive an autocorrelation function r (tau) to derive the required covariance matrix entries in the standard equations 7, )) may be combined with the composite waveform cross-correlation (W _αβ (τ)). These equations can then be solved to obtain the prediction coefficients.

)가 제한된 수의 파라미터들을 사용하여 기술될 수 있는 것일 수 있다. 자기 상관 함수(r(τ))를 기술하는 제한된 수의 파라미터들이 디코더(100)에 송신될 수 있다. 디코더(100)의 예측자 계산기(105)는 수신된 파라미터들로부터 자기 상관 함수(r(τ))를 결정할 수 있고 표준 방정식(7)이 결정될 수 있는 서브대역 신호들의 공분산 행렬(R_αβ)을 결정하기 위해 식(3)을 사용할 수 있다. 표준 방정식(7)은 이후 예측자 계산기(105)에 의해 풀이될 수 있고, 그에 의해 예측 계수들(c_β)을 산출한다.In other words, the input audio signal to be encoded can be modeled by a process (x (t)) which may be described using a limited number of model parameters. In particular, the modeling process x (t)

) May be described using a limited number of parameters. A limited number of parameters describing the autocorrelation function r ([tau]) may be transmitted to the decoder 100. [ The predictor calculator 105 of the decoder 100 can determine the autocorrelation function r (tau) from the received parameters and determine the covariance matrix R _{alpha beta} of the subband signals for which the standard equation 7 can be determined Equation (3) can be used to determine. The standard equation (7) can then be solved by the predictor calculator 105, thereby calculating the prediction coefficients c _?.

Next, exemplary signal models that may be used to apply the model-based prediction scheme described above in an efficient manner are described. The signal models described below are generally largely related to, for example, coding rate signals for coded audio signals.

One example of a signal model is a sinusoidal process

, And the random variables (a, b) are not correlated, and have a zero mean, and variance 1. The autocorrelation function of this sinusoidal process is given by:

)을 포함하는 다중 사인 모델이다:The generalization of this sinusoidal process involves the use of a set of (angular) frequencies S, i.e. a plurality of different (angular) frequencies

): &Lt; RTI ID = 0.0 &gt;

,

)이 쌍으로 상관되지 않는다면, 제로 평균 및 분산 1을 갖고, 다중 사인 프로세스는 자기 상관 함수를 갖는다:All random variables (

,

) Are not correlated with each other, the zero mean and variance 1, and the multiple sign process has autocorrelation function:

The power spectral density (PSD) of the multinomial process (corresponding to the Fourier transform of the autocorrelation function) is the line spectrum:

)를 갖는 완화된 다중 사인 프로세스로의 교체를 초래할 수 있고,

는 비교적 작은 완화 파라미터이다. 후자의 모델은 임펄스 함수들이 없는 완전히 양의 PSD를 초래한다.Numerical observations show that a pure multinomial process with an autocorrelation function of a mathematical process is called an autocorrelation function

&Lt; RTI ID = 0.0 &gt; and / or &lt; / RTI &

Is a relatively small relaxation parameter. The latter model results in a completely positive PSD without impulse functions.

Examples of simple descriptions of the set of frequencies (S) of the multi-sine model are:

1. One fundamental frequency Ω:

2. M fundamental frequencies: Ω ₀ , Ω ₁ , ..., Ω _M-1 :

3. One sideband shifted fundamental frequency Ω, θ:

, a는 모델의 비조화 성분을 기술한다.4. Weak harmonic model: Ω, a:

, a describes the non-harmonics of the model.

As such, a (possibly relaxed) multisignal model representing the PSD given by equation (12) can be described in an efficient manner using one of the exemplary techniques listed above. By way of example, a complete set S of frequencies in the line spectrum of equation (12) can be described using only one fundamental frequency, [Omega]. Regardless of the number of prediction coefficients used by the subband predictor 103 (that is, the prediction), the input audio signal to be encoded can be well described using a multi-sine model that represents one fundamental frequency The model-based predictor is described by one parameter (i.e., by the fundamental frequency, [Omega]), independent of the masks 202, 203, 204,

1 for describing a set of frequencies S yields a process x (t) for modeling input audio signals having a period (T = 2 pi / [Omega]). When experiencing the inclusion of a zero frequency (DC) contribution with a variance of 1/2 for Eq. (11) and the resizing of the result by a factor (2 / T), the autocorrelation of the periodic model process (x The function can be written as:

)의 정의에 의해, 주기적 모델의 완화 버전의 자기 상관 함수는 다음과 같이 주어진다:Mitigation Factor (

), The autocorrelation function of the relaxed version of the periodic model is given by:

Equation (14) can also be applied to the process defined by one delay loop to which the white noise z (t) is supplied, i.e., the autocorrelation function of the model process

.

This means that a periodic process representing one fundamental frequency (?) Corresponds to a delay in the time domain, and the delay is T = 2? /?.

,

)의 단위 분산 가정에 의해 일반적으로 평탄한 대규모 전력 스펙트럼을 갖는다. 그러나, 신호 모델들은 일반적으로 임계적으로 샘플링된 필터뱅크의 서브대역들의 서브세트에 대해 로컬로만 고려된다는 것이 주의되어야 하고, 필터뱅크는 전체 스펙트럼의 성형시 중요하다. 다시 말해서, 서브대역 폭들과 비교하여 느린 변동을 갖는 스펙트럼 형상을 갖는 신호에 대하여, 평탄한 전력 스펙트럼 모델들은 신호에 양호한 매칭을 제공할 것이고, 이후, 모델 기반 예측자들은 예측 이득의 적절한 레벨들을 제공할 것이다.The above-mentioned general signal models include sinusoidal amplitude parameters (

,

Lt; / RTI &gt; has a generally flat, large-scale power spectrum. It should be noted, however, that the signal models are generally only considered locally for a subset of the subbands of the filter bank that are critically sampled, and the filter bank is important in shaping the entire spectrum. In other words, for a signal having a spectral shape with slow variations compared to subband widths, the flat power spectral models will provide a good match to the signal, and then the model-based predictors will provide appropriate levels of prediction gain will be.

More generally, the PSD model can be described by standard parameterizations of automatic regression (AR) or automatic session moving average (ARMA) processes. This increases the performance of the model-based prediction at a possible cost of increase in the technical model parameters.

)로 된다. 예를 들면, 관련된 비정상 사인파 모델들은 진폭(AM) 및 주파수 변조(FM)를 포함할 수 있다.Other variations are obtained by discarding the normality hypothesis for a stochastic signal model. The autocorrelation function is then used to determine the function of the two variables

). For example, the associated abnormal sinusoidal models may include amplitude (AM) and frequency modulation (FM).

In addition, a more deterministic signal model may be employed. As will be appreciated in some of the examples below, prediction may have an extinction error in some cases. In these cases, the probabilistic approach can be avoided. When the prediction is complete for all signals in the model space, it is not necessary to perform the prediction of the mean value by the probability measurement on the considered model space.

Next, a number of aspects relating to modulated filter banks are described. In particular, aspects are described that affect the determination of the covariance matrix, thereby providing an efficient way to determine the prediction coefficients of the subband predictor. The modulated filter bank may be any combination of synthetic waveforms; (N, k), where n = 0, 1, ... is a subband index (frequency band) and k ∈ z is a subband sample index )to be. For ease of explanation, the composite waveforms are given in consecutive times and the unit time width

, Where in the case of the cosine modulation filter bank: &lt; RTI ID = 0.0 &gt;

It is assumed that the window function (v (t)) is a real number and is an even number. Up to small fluctuations of the modulation rule, it has L subbands at sampling at time step (1 / L), modulated discrete cosine transform (MDCT), quadrature phase mirror filter (QMF) ). &Lt; / RTI &gt; It is assumed that the window is of length and finite duration with the support contained in the interval [-K / 2, K / 2], K is the overlap factor of the nested transform and K is the length of the window function.

By the shift invariant structure, we found that the cross-correlation function of the composite waveform (as defined in equation (4)) can be written as:

이고, 정의

를 갖는다. 변조 구조(17)는,In other words,

And definition

. The modulation structure (17)

)는 필터뱅크 윈도우의 와그너-빌 분포의 주파수 변수에서 필터뱅크 서브대역 단차를 갖는 샘플링을 나타낸다:For other extensions. Kernel function (

) Represents the sampling with the filter bank subband step in the frequency variable of the Waveness-Bill distribution of the filter bank window:

The kernel is real and even in both v and tau, by the above-mentioned assumptions for the window function (v (t)). Its Fourier transform is the product of the shifted window responses,

)은 |τ| > K에 대해 0이 되고 필터뱅크 윈도우들(ν(t))의 일반적인 선택들에 대해 |ν|의 함수로서 빠른 감쇠를 갖는 식들(20, 21)로부터 이해될 수 있다. 결과로서, ν = n + m + 1을 포함하는 식(19)의 제 2 항은 가장 낮은 서브대역들을 제외하고 종종 무시될 수 있다.to be. This is the kernel (

) Is | τ | Can be understood from equations (20, 21) with zero attenuation as a function of | v | for the general choices of filter bank windows (v (t)). As a result, the second term of equation (19), including v = n + m + 1, can often be ignored except for the lowest subbands.

For the autocorrelation function r (?) Of a given signal model, the above-mentioned equations can be inserted into the definition of the subband sample covariance matrix given by equation (3). Justice,

를 얻는다.By

.

As a function of the power spectral density (P (?)) Of a given signal model (corresponding to the Fourier transform of the autocorrelation function r (?)), The following was found:

는

의 푸리에 변환이고, n, m은 서브대역 인덱스들을 식별하고, λ는 타임 슬롯 래그(λ = k - l)를 나타낸다. 식(23)의 수식은 다음과 같이 다시 쓸 수 있다:

The

N, m identifies subband indices, and lambda represents a timeslot lag (lambda = k - l). The formula in Eq. (23) can be rewritten as:

)에서 대응하는 시프트를 찾고, 부호는 시간 래그(λ)의 (정수) 값들에 의존한다. 이는, 일반적인 필터 뱅크 경우와 비교하여, 변조 구조를 갖는 필터뱅크를 사용하는 이점을 반영한다.An important observation is that the first term in equation (24) necessarily has an invariant property with respect to frequency shifts. When the second term of equation (24) is shifted to P (? -? V) by the product of integer (?) And subband spacing (?),

) And the sign depends on the (integer) values of the time lag lambda. This reflects the advantage of using a filter bank with a modulation structure, as compared to the case of a general filter bank.

)에서 하나의 사인파를 포함하는 신호 모델(x(t))을 사용하는 사인파 모델 기반 예측 방식의 경우, PSD는

에 의해 주어진다. P(ω)를 수식(24)에 삽입하는 것은 네 개의 항들을 제공하고, 그 중 세 개는 n + m + 1이 크다는 가정 하에서 세 개가 무시될 수 있다. 나머지 항은,Equation (24) provides an efficient means for determining the matrix coefficients of the subband sample covariance matrix when knowing the PSD of the underlying signal model. By way of example, the (angular) frequency (

In the case of a sinusoidal model-based prediction scheme that uses a signal model x (t) that includes one sinusoidal wave in the PSD,

Lt; / RTI &gt; Inserting P (?) Into equation (24) provides four terms, three of which can be ignored under the assumption that n + m + 1 is large. The remaining terms,

)은 고려된 주파수에 의해 상당히 영향을 받는 것으로 가정되는 주위의 서브대역 샘플들의 집합

에 의해 신뢰할 수 있게 예측될 수 있다. 절대 주파수(

)는 서브대역의 중심 주파수(

)에 관련된 관련 항들에서

로서 표현될 수 있고, p는 주파수(

)를 포함하는 서브대역의 서브대역 인덱스이고, f는 -0.5와 +0.5 사이의 값들을 취하고 서브대역(p)의 중심 주파수에 관한 주파수(

)의 위치를 나타내는 정규화된 주파수 파라미터이다. 서브대역 공분산 행렬(R_{n ,m})을 결정하면, 샘플 인덱스(k)에서 서브대역(n) 내 서브대역 샘플을 추정하기 위해 샘플 인덱스(l)에서 서브대역(m) 내 서브대역 샘플에 적용되는 예측자 계수들(c_m[l])이 표준 방정식들(7)을 풀이함으로써 발견되고, 상기 경우에 대해 바로 다음과 같이 기재될 수 있다:Equation (25) provides an efficient means for determining the subband covariance matrix R _{n , m} . Subband samples (

) Is the set of surrounding subband samples that are assumed to be significantly affected by the considered frequency

Lt; / RTI &gt; Absolute frequency (

) Is the center frequency of the subband (

),

And p is the frequency (

), F is a subband index of the subband including the center frequency of the subband p, taking values between -0.5 and +0.5

) &Lt; / RTI &gt; Determining the subband covariance matrix R _{n , m} applies to subband samples in subband m in sample index 1 to estimate subband samples in subband n at sample index k (C _m [l]) are found by solving the standard equations (7), and can be described as follows:

In equation (26), set B describes, for example, the prediction mask support shown in FIG. In other words, set B identifies subbands m and sample indices 1 that are used to predict the target sample.

Next, solutions of standard equations (26) for different prediction mask supports (as shown in FIG. 2) are provided in an exemplary manner. An example of a causal second-order in-band predictor is obtained by selecting a prediction mask support B = {(p, -1), (p, -2)}. This prediction mask support corresponds to the prediction mask 202 of FIG. Using the approximation of equation (25), the standard equations (26) for these two tap predictions are:

, c_p[-2] = -1로 주어지고, 이는 주파수(

)가

이도록 선택되지 않는 한 유일하다. 식(6)에 따른 제곱된 예측 에러의 평균값이 0인 것을 찾았다. 따라서, 사인파 예측은 식(25)의 근사까지 완료한다. 주파수 시프트들에 대한 불변 속성은, 정의(

)를 사용하여, 예측 계수(c_p[-1])가 정규화된 주파수(f)에 의해

로서 다시 쓸 수 있다는 사실에 의해 여기에 예시된다. 이는, 예측 계수들이 특정 서브대역내 정규화된 주파수(f)에만 의존한다는 것을 의미한다. 그러나, 예측 계수들의 절대값들은 서브대역 인덱스(p)에 독립적이다.The solution to equation (27) is

, c _p [-2] = -1, which is the frequency

)end

Unless selected to be. We found that the mean value of the squared prediction error according to Eq. (6) is zero. Thus, the sine wave prediction is completed up to the approximation of equation (25). The invariant properties for frequency shifts are defined by (

), The prediction coefficient c _p [-1] is multiplied by the normalized frequency f

As shown in FIG. This means that the prediction coefficients depend only on the normalized frequency f in a particular subband. However, the absolute values of the prediction coefficients are independent of the subband index p.

As discussed above with respect to FIG. 4, in-band prediction has certain disadvantages with aliasing artifacts in noise shaping. The following example relates to the improved behavior shown in Fig. The causal cross-band prediction taught herein is obtained by selecting a prediction mask support B = {(p-1, -1), (p, -1), (p + 1, -1) Requires only one earlier time slot instead of two, and this performs noise shaping with fewer alleys frequency contributions than the classical prediction mask 202 of the first example. The prediction mask support B = {(p-1, -1), (p, -1), (p + 1, -1)} corresponds to the prediction mask 203 of FIG. The standard equations 26 based on the approximation of Eq. (25) are based on two equations for three unknown coefficients (c _m [-1], m = p-1, p, p + 1) Is reduced:

It has been found that any solution to equation (28) results in an average value of zero of the squared prediction error according to equation (6). A possible strategy for choosing one solution of the infinite number of solutions for equation (28) is to minimize the sum of the squares of the prediction coefficients. this is,

&Lt; / RTI &gt; It is clear from equation (29) that the prediction coefficients only depend on the normalized frequency f with respect to the midpoint of the target subband p and also on the parity of the target subband p.

)을 예측하기 위해 동일한 예측 마스크 지원(B = {(p-1, -1), (p, -1), (p+1, -1)})을 사용함으로써, 3 × 3 예측 행렬이 얻어진다. 표준 방정식들에서 모호성을 피하기 위해 더 자연스러운 전략의 도입시, 즉,

에 대응하는 완화된 사인파 모델(

)을 삽입함으로써, 수치 계산들은 도 3의 3 × 3 예측 행렬 요소들을 초래한다. 예측 행렬 요소들은 사인파 윈도우 함수(ν(t) = cos(πt/2))를 갖는 중첩(K=2)의 경우에 및 홀수 서브대역(p)의 경우에 정규화된 주파수의 함수(

)로서 보여진다.Three subband samples for m = p-1, p, p + 1 (shown as prediction mask 204 in FIG. 2

(P = 1, -1), (p, -1), (p + 1, -1)} is used to predict a prediction matrix Loses. In introducing more natural strategies to avoid ambiguity in standard equations,

A relaxed sine wave model corresponding to

), The numerical calculations result in the 3 × 3 prediction matrix elements of FIG. The prediction matrix elements are a function of the normalized frequency in the case of the superposition (K = 2) with the sinewave window function (v (t) = cos (t / 2)

).

As such, it has been shown that the signal model x (t) is used to describe the basic features of the input audio signal to be encoded. The parameters describing the autocorrelation function r (tau) can be transmitted to the decoder 100 so that the decoder 100 derives predictors from the transmitted parameters and from knowledge of the signal model x (t) Let's calculate. For modulated filter banks, it has been shown that efficient means can be derived to determine the subband covariance matrix of the signal model and to solve the standard equations to determine the predictor coefficients. In particular, it has been shown that the resultant predictor coefficients are invariant to the subband shifts and generally depend only on the normalized frequency with respect to the particular subband. As a result, there are predetermined reference tables (for example, &lt; RTI ID = 0.0 &gt; e. G. &Lt; / RTI &gt; As shown in FIG. 3) may be provided.

Next, using a single fundamental frequency (?), For example, periodic model-based prediction is described in more detail. The autocorrelation function r (τ) of this periodic model is given by Eq. (13). An even PSD or line spectrum is given by:

)에 가장 가까운 부분 주파수(

)를 사용하여 상기에 도출되는 사인파 모델의 적용을 허용하기에 충분히 크다. 이는 작은 주기(T), 즉, 필터뱅크의 시간 스트라이드에 관하여 작은 기간을 갖는 주기적 신호들이 상기 기술된 사인파 모델을 사용하여 양호하게 모델링 및 예측될 수 있다는 것을 의미한다.When the period T of the periodic model is sufficiently small, for example T? 1, the fundamental frequency? = 2? / T is the center frequency of the subband p of the target subband sample to be predicted

) To the nearest partial frequency (

) To allow the application of the sine wave model derived therefrom. This means that periodic signals with a small period (T), i.e. a small period with respect to the time stride of the filter bank, can be well modeled and predicted using the sine wave model described above.

When the period T is sufficiently large compared to the duration K of the filter bank window v (t), the predictor can be reduced to an approximation of the delay by T. [ As can be seen, these predictor coefficients can be read directly from the waveform cross-correlation function given by equation (19).

Insertion of the model according to equation (13) into equation (22) results in:

를 따른다. 대체로, 표준 방정식들(7)은 다음과 같이 된다:An important observation is that for T ≥ 2K, at most one term in Eq. (31) is not 0 for each λ, since U _{n, m} (τ) = 0 for | τ |> K. (N, k), (m, l) E B implies | kl | TK by selecting a prediction mask support (B = I x J) with a time slot diameter (D = | J | , And therefore the single term in equation (31) is for q = 0. This is the inner product of orthogonal waveforms, and is zero unless n = m and k = l

. In general, the standard equations (7) are as follows:

주위에 중심을 두도록 선택될 수 있고, 이 경우에 식(32)의 우변은 q = -1로부터의 그의 단일 기여를 갖는다. 이후, 계수들은 다음으로 주어지고:Predictive mask support

, And in this case the right hand side of equation (32) has its single contribution from q = -1. The coefficients are then given as: &lt; RTI ID = 0.0 &gt;

)의 보완에 의해 걸쳐진 공간상에 u_p(t+T)의 투영의 제곱 평균과 같다.An explicit expression from Eq. (19) can be inserted. The geometry of the prediction mask support for this case may have the appearance of a prediction mask support of the prediction mask 205 of FIG. The average value of the squared prediction error given by equation (6) is the approximate waveforms w _{m , l} (t)

) Of the projection of u _p (t + T) onto the space spanned by the complement of u _p (t + T).

)(서브대역(p)으로부터 및 시간 인덱스(0)에서) T와 대략 같은 시간 다이어미터를 갖는 (p, -T) 주위에 중심을 둔 적절한 예측 마스크 지원(B)을 사용함으로써 예측될 수 있다. 표준 방정식들은 T 및 p의 각각의 값에 대해 풀이될 수 있다. 다시 말해서, 입력 오디오 신호의 각각의 주기성(T)에 대하여 및 각각의 서브대역(p)에 대하여, 주어진 예측 마스크 지원(B)에 대한 예측 계수들은 표준 방정식들(33)을 사용하여 결정될 수 있다.In view of the above, the subband samples (

May be predicted by using the appropriate prediction mask support B centered around (p, -T) with a time diagram approximately equal to T (at sub-band p and at time index 0) . Standard equations can be solved for each value of T and p. In other words, for each periodicity T of the input audio signal and for each subband p, the prediction coefficients for a given prediction mask support B can be determined using the standard equations 33 .

, q∈Z) 중에서 양의 주파수들의 서브세트(S(θ))에 의해 규정된, 서브대역(p)의 중심 주위에 중심을 둔, 즉,

주위에 중심을 둔, 시프트 파라미터(-1/2≤θ≤1/2)를 갖는 시프트된 조화 모델을 연구하기 위해 일반적으로 충분할 것이다:Due to the large number of subbands p and the wide range of periods T, the creation of a direct table of all predictor coefficients is not practical. However, in a manner similar to a sinusoidal model, the modulation scheme of the filter bank provides a significant reduction in the required table size through an invariant property on the frequency shifts. This means that the set of frequencies

, Centered around the center of the sub-band (p) defined by a subset (S (θ)) of the positive frequency from q∈ Z), that is,

Will typically suffice to study a shifted harmonic model with a shift parameter centered around (-1 / 2 &amp;le;&amp;le;&amp;

In fact, considering a T and a sufficiently large subband index p, the cyclic model according to equation (30) can be recovered to a good approximation by the shifted model according to equation (34) by appropriate selection of the shift parameter [ . (N and p define the subband indices around the subband p of the prediction mask support) and the Fourier analysis &lt; RTI ID = 0.0 &gt; Based adjustments result in the following equation for the covariance matrix:

As can be appreciated, equation (35) depends on the target only through the factor ((-1) ^pλ) subband index (p). For the case of the large period T and the small temporal lag λ, we have found that the term for l = 0 only contributes to equation (35) and again the covariance matrix is the unit matrix. The right side of the standard equations (26) for the appropriate prediction mask support (B) centered around (p, -T) then directly provides the prediction coefficients as follows:

)의 정규 선택에 의해 식(19) 내지 식(33)의 제 1 항의 기여를 복구한다.That is,

(19) to (33) by the normal selection of the second term (i. E.

Equation (36) allows to determine the prediction coefficients c _{p + v} [k] for the subband p + v at the time index k, p). As can be seen from equation (36), the prediction coefficients c _{p + v} [k] can be obtained only from the target subband index p through the factor (-1) ^pk that affects the sign of the prediction coefficient, Lt; / RTI &gt; However, the absolute value of the prediction coefficient is independent of the target subband index p. On the other hand, the prediction coefficient c _{p + v} [k] depends on the periodicity T and the shift parameter?. Further, the prediction coefficient (c _{p + v} [k]) depends on v and k, i.e., the prediction mask support (B) used to predict the target sample in the target subband (p).

)를 요구하고, 그래서 시프트 파라미터(θ)에 대한 양자화 단차 크기는 주기성(T)에 기초하여 크기 조정될 것이다.In the present specification, it is proposed to provide a look-up table that allows searching for a set of prediction coefficients (c _{p + v} [k]) for a predetermined prediction mask support (B). For a given prediction mask support B, the lookup table provides a set of prediction coefficients c _{p + v} [k] for a predetermined set of values of the periodicity T and the shift parameter [theta] do. In order to limit the number of reference table entries, the number of predetermined values of the periodicity (T) and the number of predetermined values of the shift parameter (?) Should be limited. As can be appreciated from equation (36), the appropriate quantization step size for the predetermined values of the periodicity (T) and the shift parameter (?) Will depend on the periodicity (T). In particular, it can be appreciated that relatively large quantization steps for the relatively large periodicity T (with respect to the duration K of the window function), for the periodicity T and for the shift parameter [theta] can be used. As another extreme, for relatively small periodicity (T) that tends to be zero, only one sine wave contribution has to be considered, so the periodicity (T) loses its importance. On the other hand, the equations for the sinusoidal prediction according to equation (29) are the slowly varying normalized absolute frequency shift

), So the quantization step size for the shift parameter [theta] will be scaled based on the periodicity T.

In general, it is proposed herein to use uniform quantization of periodicity (T) with a fixed step size. The shift parameter [theta] may also be quantized with a step size proportional to min (T, A), but the value of A depends on the characteristics of the filter bank window function. Further, for T < 2, the range of the shift parameters [theta] is limited to |? |? Min (CT, 1/2) for some constant C Lt; / RTI &gt;

로 치환에 의하여 재해석될 수 있다. 다음을 찾을 수 있다:FIG. 6A shows an example of a resultant quantization grid in the (T,?) - plane for A = 2. In the intermediate range only in the range of 0.25? T? 1.5, full two-dimensional dependence is considered, whereas one-dimensional parameterization given by equations (29) and (36) can be used for the remaining range of interest have. In particular, for periodicities T that tend to be zero (e.g., T < 0.25), periodic model-based predictions substantially correspond to sinusoidal model-based predictions, and prediction coefficients use equations 29 &Lt; / RTI &gt; On the other hand, for the periodicity T (e.g., T> 1.5) that substantially exceeds the window duration K, the prediction coefficients c _{p + v} [k] using periodic model- ) May be determined using equation (36). These expressions

Can be reinterpreted by substitution. You can find:

By providing a role given by φ in the table as a parameter (θ), an essentially separable structure is obtained in the equivalent (T, φ) -plane. The dependence on T is included in the first slowly varying factor and the dependence on? Is included in the one-periodic second factor in equation (37), up to signal changes depending on the subband and time slot indices.

The modified offset parameter phi may be interpreted as a shift of the harmonic series by the units of the fundamental frequency measured from the midpoint of the midpoints of the source and target bins. Since the symmetries of the explicit equation (37) with respect to the simultaneous sign changes of phi and v are generally to be retained and can be used to reduce the table sizes, this modified parameterization (T, &amp;phiv;).

As shown above, FIG. 6A shows a two-dimensional quantization grid based on statistical tables for periodic model-based predictor computation in a cosine-modulated filter bank. The signal model is a model of a signal 602 having a period T measured in units of filter bank time steps. Equally, the model includes integer lines of frequency lines, also known as parts of the fundamental frequency corresponding to period T. For each target subband, the shift parameter (?) 601 represents the distance of the portion closest to the center frequency measured in units of fundamental frequency (?). The shift parameter (?) 601 has a value between -0.5 and 0.5. The black crosses 603 of FIG. 6A illustrate the proper density of quantization points for a table of predictors with a high prediction gain based on a periodic model. For large periods T (e.g., T > 2), the grid is even. The increased density at the shift parameter [theta] is generally required that the period T decreases. However, in the area outside the lines 604, the distance [theta] is greater than one frequency bin of the filter bank, so that most of the grid points in this area can be ignored. The polygon 605 defines a sufficient area for the entire table. In addition to the slightly outwardly sloping lines of lines 604, boundaries are introduced at T = 0.25 and T = 1.5. This means that small cycles 602 can be handled as separate sine waves and predictors for large cycles 602 can be processed in a one-dimensional (&quot;Lt; / RTI &gt; can be approximated by means of tables. For the embodiment shown in FIG. 6A, the prediction mask support is generally similar to the prediction mask 205 of FIG. 2 for large periods (T).

)에 위치된다. 필터뱅크의 서브대역들의 중심 주파수들은 주파수들(

;

)에 위치된다. 주어진 서브대역(p)에 대하여, 주어진 서브대역의 중심 주파수(

)에 가장 가까운 주파수(ω=qΩ)를 갖는 펄스(616)의 주파수 위치는 비교항들에서

로서 기술될 수 있고, 시프트 파라미터(θ)는 -0.5로부터 +0.5의 범위이다. 이와 같이, 항(

)은 중심 주파수(

)로부터 조화 모델의 가장 가까운 주파수 성분(616)까지의 거리(주파수로)를 반영한다. 이는 도 6b의 상부 도면에 도시되고 중심 주파수(617)는

이고, 거리(618)(

)는 비교적 큰 주기(T)의 경우에 대해 도시된다. 시프트 파라미터(

)가 서브대역(p)의 중심의 관점으로부터 보여지는 완전 조화 급수들을 기술하는 것을 허용한다는 것이 이해될 수 있다.FIG. 6B shows a cyclic model-based prediction in the case of relatively large periods T and in the case of relatively small periods T. FIG. For the large periods T, that is, for relatively small fundamental frequencies (OMEGA) 613, the window function 612 of the filter bank is applied to the DLEAK pulses 616 of the PSD of a relatively large number of lines or periodic signals From the top figure. Dail pulse 616 includes frequencies 610 (? = Q?;

). The center frequencies of the subbands of the filter bank are frequency

;

). For a given subband p, the center frequency of a given subband (

The frequency position of the pulse 616 having the closest frequency (? = Q?) To the frequency

And the shift parameter [theta] ranges from -0.5 to +0.5. As such,

) Is the center frequency (

(Frequency) of the harmonic model to the closest frequency component 616 of the harmonic model. This is illustrated in the top view of FIG. 6B and center frequency 617 is

, And distance 618 (

) Is shown for the case of a relatively large period T. The shift parameter (

) Allows describing the perfect harmonic series as seen from the perspective of the center of the subband p.

The lower figure of Figure 6b shows the case for relatively small periods T, i.e. for relatively large fundamental frequencies OMEGA 623, and in particular for fundamental frequencies 623 larger than the width of window 612 Respectively. It can be appreciated that in these cases, the window function 612 may include only a single pulse 626 of the periodic signal so that the signal can be viewed as a sinusoidal signal in the window 612. This means that for relatively small periods T, the periodic model-based prediction scheme converges to a sinusoidal model-based prediction scheme.

6B also shows exemplary prediction masks 611 and 621 that can be used for the periodic model-based prediction scheme and for the sinewave model-based prediction scheme, respectively. The prediction mask 611 used in the periodic model based prediction scheme may correspond to the prediction mask 205 of FIG. 2 and may include prediction mask support 614 to estimate the target subband samples 615. The prediction mask 621 used in the sinusoidal model based prediction scheme may correspond to the prediction mask 203 in FIG. 2 and may include prediction mask support 624 to estimate the target subband samples 625.

7A illustrates an exemplary encoding method 700 that includes model-based subband prediction using a periodic model (e.g., comprising one fundamental frequency, [Omega]). The frame of the input audio signal is considered. For such a frame, the periodicity T or the fundamental frequency OMEGA can be determined (step 701). The audio encoder may include elements of the decoder 100 shown in FIG. 1, and in particular, the audio encoder may include a predictor calculator 105 and a subband predictor 103. The periodicity T or the fundamental frequency OMEGA can be determined such that the mean value of the squared predicted error subband signals 111 is reduced (e.g., minimized) according to equation (6). By way of example, the audio encoder may use the different fundamental frequencies [Omega] to determine the predicted error subband signals 111 and to reduce (e.g., minimize) the average value of the squared predicted error subband signals 111 ) Can be applied to determine the fundamental frequency (?). The method proceeds to quantizing the resulting prediction error subband signals 111 (step 702). The method also includes generating (703) a bitstream that includes information indicating the determined fundamental frequency ([Omega]) and the quantized predicted error subband signals (111).

When determining the fundamental frequency [Omega] in step 701, the audio encoder may use the equations (36) and / or (29) to determine the prediction coefficients for a particular fundamental frequency [Omega]. The set of possible fundamental frequencies ([Omega]) may be limited to the number of bits available for transmission of information indicating the determined fundamental frequency ([Omega]).

The audio coding system may comprise a predetermined model (e.g., a periodic model comprising one fundamental frequency (?) Or any other of the models provided herein) and / or a predetermined prediction mask 202, 203, 204 , 205) may be used. On the other hand, the audio coding system may be provided with additional degrees of freedom by allowing the audio decoder to determine an appropriate prediction mask and / or an appropriate model for the audio signal to be encoded. Information about the selected model and / or the selected prediction mask is then encoded into a bitstream and provided to a corresponding decoder 100.

FIG. 7B illustrates an exemplary method 710 for decoding an encoded audio signal using model-based prediction. It is assumed that the decoder 100 knows the signal model and predictive mask used by the encoder (either via the received bitstream or by predetermined settings). It is also assumed that a cyclic prediction model is used for exemplary purposes. The decoder 100 extracts information on the fundamental frequency [Omega] from the received bitstream (step 711). Using information about the fundamental frequency ([Omega]), the decoder 100 can determine the periodicity (T). The fundamental frequency ([Omega]) and / or the periodicity (T) may be used to determine a set of prediction coefficients for different subband predictors (step 712). The subband predictors may be used to determine the estimated subband signals that are combined with the semi-quantized predicted error subband signals 111 (step 714) to yield decoded subband signals 113 (Step 713). The decoded subband signals 113 may be filtered using the synthesis filter bank 102 (step 715), thereby yielding the decoded time domain audio signal 114. [

)를 포함하는 것으로 결정된다. 이후, 정규화된 주파수(f)는 관계(

)를 사용하여 결정되고, 주파수(

)는 서브대역(p)에 있는 곱(ω=qΩ)에 대응한다. 예측자 계산기(105)는 이후 (예를 들면, 도 2의 예측 마스크(203) 또는 도 6b의 예측 마스크(621)를 사용하여) 예측 계수들의 세트를 결정하기 위해 식(29) 또는 미리 결정된 참조표를 사용할 수 있다.The predictor calculator 105 may use equations (36) and / or (29) to determine the predictive coefficients of the subband predictors 103 based on the received information about the fundamental frequency (Step 712). This can be done in an efficient manner using reference tables as shown in Figures 6A and 3. [ By way of example, the predictor calculator 105 may determine the periodicity T and determine whether the periodicity is below a predetermined lower threshold (e.g., T = 0.25). In this case, a sine wave model based prediction method is used. This is because, based on the received fundamental frequency ([Omega]), the subbands p are multiplied by a multiple of the fundamental frequency ([omega] = q [

). Then, the normalized frequency f is expressed by the relationship (

), And the frequency (

) Corresponds to the product (? = Q?) In subband p. The predictor calculator 105 may then use Equation 29 or a predetermined reference to determine the set of prediction coefficients (e.g., using the prediction mask 203 of Figure 2 or the prediction mask 621 of Figure 6b) You can use tables.

)에 의해 상당히 영향을 받는 서브대역들(p)에 대해서만 결정된다. 다른 서브 대역들에 대해, 이러한 다른 서브대역들에 대한 추정된 서브대역 신호들(112)이 0인 것을 의미하는 예측 계수들이 결정되지 않는다. 디코더(100)의(및 동일한 예측자 계산기(105)를 사용하여 인코더의) 계산 복잡성을 감소시키기 위하여, 예측자 계산기(105)는 T 및

에 대한 값들을 조건으로 하는 예측 계수들의 세트를 제공하는 미리 결정된 참조표를 이용할 수 있다. 특히, 예측자 계산기(105)는 T에 대하여 복수의 상이한 값들에 대한 복수의 참조표들을 이용할 수 있다. 복수의 참조표들의 각각은 시프트 파라미터(

)의 복수의 상이한 값들에 대한 예측 계수들의 상이한 세트를 제공한다.It should be noted that a different set of prediction coefficients may be determined for each subband. However, in the case of a sinusoidal model-based prediction scheme, a set of prediction coefficients is generally a multiple of the fundamental frequency (? = Q?

Lt; RTI ID = 0.0 &gt; p &lt; / RTI &gt; For the other subbands, the prediction coefficients that mean that the estimated subband signals 112 for these other subbands are zero are not determined. In order to reduce the computational complexity of the decoder 100 (and of the encoder using the same predictor calculator 105), the predictor calculator 105 calculates T and /

Lt; / RTI &gt; may use a predetermined look-up table that provides a set of prediction coefficients that are conditional on the values for &lt; RTI ID = 0.0 &gt; In particular, the predictor calculator 105 may use a plurality of look-up tables for a plurality of different values for T. Each of the plurality of look-up tables includes a shift parameter

&Lt; / RTI &gt; for a plurality of different values of the prediction coefficients.

)의 함수로서 또는 변경된 시프트 파라미터(φ)의 함수로서 제공할 수 있다. T=8/32 내지 T=80/32까지에 대한 참조표들은 단차 크기의 절반씩 증가되는 범위, 즉,

에 대해 사용될 수 있다. 참조표들이 정의된 이용가능한 주기성들과 상이한 주어진 주기성에 대하여, 가장 근접한 이용가능한 주기성에 대한 참조표가 사용될 수 있다. 상기 개요로 서술된 바와 같이, 긴 주기들(T)에 대하여(예를 들면, 참조표가 정의되는 주기를 초과한 주기들(T)에 대하여), 식(36)이 사용될 수 있다. 대안적으로, 참조표들이 정의된 주기들을 초과하는 기간들(T)에 대하여, 예를 들면, 주기들(T>81/32)에 대하여, 주기(T)는 정수 지연(T_i) 및 나머지 지연(T_r)으로 분리될 수 있고, T = T_i + T_r이다. 상기 분리는 나머지 지연(T_r)이 식(36)이 적용가능하고 참조표들이 이용가능한 간격 내에, 예를 들면, 상기 예에 대하여 간격([1.5, 2.5] 또는 [49/32, 81/32]) 내에 있는 것일 수 있다. 이렇게 함으로써, 예측 계수들은 나머지 지연(T_r)에 대하여 참조표를 사용하여 결정될 수 있고, 서브대역 예측자(103)는 정수 지연(T_i)에 의해 지연된 서브대역 버퍼(104)상에 동작할 수 있다. 예를 들면, 주기가 T=3.7인 경우, 정수 지연은 T_i = 2이고, 다음으로 나머지 지연은 T_r = 1.7이다. 예측자는 (추가의) T_i = 2에 의해 지연되는 신호 버퍼상의 T_r = 1.7에 대한 계수들에 기초하여 적용될 수 있다.In an actual implementation, a plurality of look-up tables may be provided for different values of the period parameter T. [ By way of example, reference tables may be provided for values of T in the range of 0.25 and 2.5 (as shown in FIG. 6A). The look-up tables may be provided for a predetermined granularity or step size of different period parameters (T). In one exemplary implementation, the step size for the normalized period parameter T is 1/16, and different lookup tables for the quantized prediction coefficients are provided for T = 8/32 to T = 80/32 . Thus, a total of 37 different lookup tables may be used to transform the quantized prediction coefficients into a shift parameter

) Or as a function of the modified shift parameter ([phi]). The reference tables for T = 8/32 to T = 80/32 are in the range increased by half of the step size,

Lt; / RTI > For a given periodicity where the reference tables differ from the defined available periodicity, a look-up table for the closest available periodicity may be used. As described above, equation (36) may be used for long periods T (e.g., for periods T exceeding the period in which the reference table is defined). Alternatively, for periods T in which reference tables exceed defined periods, for example, for periods (T &gt; 81/32), period T may be an integral delay (T _i ) Can be separated into a delay (T _r ), and T = T _i + T _r . The separation is such that the remainder of the delay (T _r ) is within the available interval of the reference table, for example, [1.5, 2.5] or [49/32, 81/32 ]). By doing so, the prediction coefficients can be determined using the look-up table for the remaining delay T _r and the sub-band predictor 103 operates on the sub-band buffer 104 delayed by the integer delay T _i . For example, if the period is T = 3.7, the integer delay is T _i = 2, and the remaining delay is then T _r = 1.7. The predictor can be applied based on the coefficients for T _r = 1.7 on the signal buffer, delayed by (additional) T _i = 2.

The separation scheme depends on a reasonable assumption that the extractor approximates the delay by T in the range [1.5, 2.5] or [49/32, 81/32]. The advantage of the separation procedure in comparison to the use of equation (36) is that the prediction coefficients can be determined based on computationally efficient table retrieval operations.

의 샘플링 단차 크기를 갖고 범위 |φ|≤T로 한정되는 것이 관찰된다.As briefly described above, for short periods (T < 0.25), equation (29) can be used to determine the prediction coefficients. Alternatively, it may be advantageous to use (readily available) reference tables (to reduce computational complexity). When the modified shift parameter phi (for T < 0.25 and C = 1, A = 1/2)

Is limited to the range | φ | ≤T.

)로 문의될 수 있다. 이를 행함으로써, 짧은 주기들(예를 들면, T<0.25)에 대한 예측 계수들은 또한 표 검색 동작들을 사용하여 계산 효율적인 방식으로 결정될 수 있다. 또한, 참조표들의 수가 감소될 수 있기 때문에, 예측자에 대한 메모리 요구 사항들은 감소될 수 있다.In the present specification it is proposed to reuse the look-up table for the lowest period T = 0.25 by resizing the modified shift parameter φ with T _l / T, where T _l is the least frequent period For example, _Tl = 0.25). For example, with T = 0.1 and phi = 0.07, the table for T = 0.25 shows the resized shift parameter (

). By doing this, prediction coefficients for short periods (e.g., T < 0.25) can also be determined in a computationally efficient manner using table retrieval operations. Also, since the number of lookup tables can be reduced, the memory requirements for the predictor can be reduced.

In this specification, a model-based subband prediction scheme has been described. The model-based subband prediction scheme enables efficient description of subband predictors, i.e., a technique requiring only a relatively small number of bits. As a result of efficient techniques for subband predictors, cross-subband prediction schemes can be used resulting in reduced aliasing artifacts. Overall, this allows the provision of low bit rate audio coders using subband prediction.

101: dequantizer 102: synthesis filter bank
103: Subband predictor 104: Subband buffer
105: Forecaster Calculator

Claims (35)

A method for estimating a first sample (615) of a first subband signal in a first subband of an audio signal, the first subband signal of the audio signal comprising a plurality of subbands of a plurality of subbands Wherein the analysis filter bank is determined using an analysis filter bank (612) comprising a plurality of analysis filters each providing a plurality of band signals,
Determining a model parameter (613) of the signal model;
(614) of the first decoded subband signal derived from the first subband signal based on the model parameter (613) based on the signal model and based on the analysis filter bank (612) Determining a prediction coefficient to be applied to the first sample (614), wherein the time slot of the previous sample (614) is before a time slot of the first sample (615); And
Determining an estimate of the first sample (615) by applying the prediction coefficient to the previous sample (614)
Wherein determining the prediction coefficients comprises determining the prediction coefficients using a reference table or an analysis function,
Wherein the lookup table or the analysis function provides the prediction coefficient as a function of a parameter derived from the model parameter,
Wherein the reference table or the analytic function is predetermined based on the signal model and based on the analysis filter bank. The method according to claim 1,
Wherein the signal model comprises one or more sine wave model components,
The model parameter 613 represents the frequency of the one or more sinusoidal model components, and
Optionally,
The model parameter 613 represents the fundamental frequency (?) Of the multi-sinusoidal signal model,
Wherein the multi-sinusoidal signal model comprises a periodic signal component,
Wherein the periodic signal component comprises a plurality of sinusoidal components,
Wherein the plurality of sinusoidal components have a frequency that is a multiple of the fundamental frequency ([Omega]), in the first subband of the audio signal. 3. The method according to claim 1 or 2,
Wherein determining the model parameter (613) comprises extracting the model parameter (613) and the model parameter (613) from a received bit stream representing a prediction error signal. A method for estimating a first sample of a first subband signal. 3. The method according to claim 1 or 2,
Wherein determining the model parameter (613) comprises determining the model parameter (613) such that an average value of the squared prediction error signal is reduced,
The prediction error signal is determined based on a difference between the estimate of the first sample 615 and the first sample 615,
Optionally,
Wherein the average value of the squared prediction error signal is determined based on a plurality of successive first samples of the first subband signal, estimating a first sample of the first subband signal in a first subband of the audio signal, Lt; / RTI &gt; 3. The method according to claim 1 or 2,
The model parameter represents a fundamental frequency (?) Of a multi-sinusoidal signal model,
Wherein determining the prediction coefficient comprises determining a multiple of the fundamental frequency (OMEGA) lying within the first subband, estimating a first sample of the first subband signal in a first subband of the audio signal, Lt; / RTI &gt; 6. The method of claim 5,
Wherein the step of determining the prediction coefficients comprises:
Selecting one of the plurality of look-up tables based on the model parameter; And
And determining the prediction coefficients based on a selected one of the plurality of look-up tables. &Lt; Desc / Clms Page number 21 &gt; The method according to claim 6,
The model parameter representing a periodicity (T);
The plurality of look-up tables include look-up tables for different values of periodicity (T);
The method comprising the step of determining the selected reference table as the reference table for the periodicity (T) indicated by the model parameter, and
Optionally,
The plurality of look-up tables include look-up tables for different values of periodicity (T) within a range of [T _min , T _max ] at a predetermined step size? T;
T _min is that for T < T _min , the audio signal can be modeled using a signal model comprising one sinusoidal model component;
The T _max is, T> with respect to the T _max, see table for the periodicity of (T _max to T _max +1) to be corresponding to the reference table for the periodicity (T _max to T _max -1) , And estimating a first sample of the first subband signal in a first subband of the audio signal. 3. The method according to claim 1 or 2,
The plurality of subbands having the same subband spacing;
The first subband is one of the plurality of subbands;
Wherein the analysis filters of the analysis filter bank are shift-invariant with respect to each other;
Wherein the analysis filters of the analysis filter bank include a common window function;
Wherein the analysis filters of the analysis filter bank include different modulated versions of the common window function;
The common window function being modulated using a cosine function;
The common window function having a finite duration (K);
The analysis filters of the analysis filter bank forming an orthogonal basis;
The analysis filters of the analysis filter bank forming a normal orthogonal basis;
The analysis filter bank comprising a cosine modulated filter bank;
Said analysis filter bank being a threshold sampled filter bank;
The analysis filter bank comprising a nested transform;
Wherein the analysis filter bank comprises one or more of MDCT, QMF, and ELT transformations; And
The analysis filter bank comprising a modulation structure; Band signal in the first sub-band of the audio signal, wherein at least one of the first sub-band signal and the second sub-band signal is true. A method for estimating a first sample of a first subband signal in a first subband of an audio signal, the first subband signal of the audio signal comprising a plurality of subband signals of a plurality of subband Wherein the analysis filter bank is determined using an analysis filter bank comprising a plurality of analysis filters each providing a plurality of analysis filters,
Determining a prediction mask (203, 205) indicative of a plurality of previous samples of a plurality of prediction mask support subbands, wherein the plurality of prediction mask support subbands comprises one of the plurality of subbands different from the first subband Determining the prediction mask (203, 205), the prediction mask comprising at least one;
Determining a plurality of prediction coefficients to be applied to the plurality of previous samples; And
Estimating a first sample of a first subband signal in a first subband of the audio signal, and determining an estimate of the first sample by applying each of the plurality of prediction coefficients to the plurality of previous samples, Way. 10. The method of claim 9,
Wherein the plurality of prediction mask support subbands are:
The first sub-band;
At least one of the plurality of subbands directly adjacent to the first subband;
Wherein the first subband includes one or more of the plurality of subbands directly adjacent to the first subband and the first subband. A method for encoding an audio signal,
Determining a plurality of subband signals from the audio signal using an analysis filter bank comprising a plurality of analysis filters;
16. A method comprising: estimating samples of the plurality of subband signals using the method of any one of claims 1, 2, 9, or 10, thereby yielding a plurality of estimated subband signals;
Determining samples of a plurality of prediction error subband signals based on corresponding samples of the plurality of subband signals and samples of the plurality of estimated subband signals;
Quantizing the plurality of prediction error subband signals; And
Generating an encoded audio signal representative of one or more parameters used to estimate the plurality of quantized prediction error subband signals and the samples of the plurality of estimated subband signals, Lt; / RTI &gt; A method for decoding an encoded audio signal, the encoded audio signal comprising a plurality of quantized prediction error subband signals and one or more parameters to be used to estimate samples of the plurality of estimated subband signals, A method for decoding, comprising:
Semi-quantizing the plurality of quantized prediction error subband signals to produce a plurality of semi-quantized prediction error subband signals;
Estimating samples of the plurality of estimated subband signals using the method of any one of claims 1, 2, 9, or 10;
Determining samples of a plurality of decoded subband signals based on corresponding samples of the plurality of estimated subband signals and samples of the plurality of semi-quantized prediction error subband signals; And
And determining a decoded audio signal from the plurality of decoded subband signals using a synthesis filter bank comprising the plurality of synthesis filters. A system (103, 105) configured to estimate one or more first samples of a first subband signal of an audio signal, the first subband signal of the audio signal providing a plurality of subband signals The system (103, 105) as defined in claim 1, wherein the system is determined using an analysis filter bank comprising a plurality of analysis filters,
A predictor calculator (105) configured to determine a model parameter of a signal model and to determine one or more prediction coefficients to be applied to one or more previous samples of a first decoded subband signal derived from the first subband signal, Wherein the one or more prediction coefficients are determined based on the model parameters and based on the analysis filter bank based on the signal model and wherein the timeslots of the one or more previous samples are time slots of the one or more first samples The predictor calculator 105; And
And a subband predictor (103) configured to determine an estimate of the one or more first samples by applying the one or more prediction coefficients to the one or more previous samples,
Wherein determining the prediction coefficients comprises determining the prediction coefficients using a reference table or an analysis function,
Wherein the lookup table or the analysis function provides the prediction coefficient as a function of a parameter derived from the model parameter,
Wherein the reference table or the analysis function is configured to estimate one or more first samples of a first subband signal of an audio signal that is predetermined based on the signal model and based on the analysis filter bank. A system (103, 105) configured to estimate one or more first samples of a first subband signal of an audio signal, the first subband signal corresponding to a first subband and the first subband signal comprising a plurality The system (103, 105) being determined using an analysis filter bank comprising a plurality of analysis filters each of which provides a plurality of subband signals in subbands, the analysis filterbank being a critically sampled filter bank, In this case,
A predictor calculator (105) configured to determine a prediction mask (203, 205) representative of a plurality of previous samples in a plurality of prediction mask support subbands, the plurality of prediction mask support subbands being different from the first subband Wherein the predictor calculator (105) is further configured to determine a plurality of prediction coefficients to be applied to the plurality of previous samples, the predictor calculator (105) comprising: at least one of the plurality of subbands; And
And a subband predictor (103) configured to determine an estimate of the one or more first samples by applying the plurality of prediction coefficients to the plurality of previous samples, respectively, And to estimate the first samples. An audio encoder configured to encode an audio signal,
An analysis filter bank configured to determine a plurality of subband signals from the audio signal using a plurality of analysis filters;
A system (103, 105) according to claim 13 or 14 configured to estimate samples of the plurality of subband signals to produce a plurality of estimated subband signals (112);
A difference unit configured to determine samples of a plurality of prediction error subband signals based on the plurality of subband signals and corresponding samples of the plurality of estimated subband signals 112;
A quantization unit configured to quantize the plurality of prediction error subband signals; And
A bitstream generation unit configured to generate an encoded audio signal representing one or more parameters used to estimate the plurality of quantized prediction error subband signals and samples of the plurality of estimated subband signals, Encoder. An audio decoder (100) configured to decode an encoded audio signal, the encoded audio signal comprising one or more parameters used to estimate a plurality of quantized prediction error subband signals and samples of a plurality of estimated subband signals , The audio decoder (100) comprising:
An inverse quantizer (101) configured to semi-quantize the plurality of quantized prediction error subband signals to produce a plurality of semi-quantized prediction error subband signals (111);
A system (105,103) according to claim 13 or 14 configured to estimate samples of the plurality of estimated subband signals (112);
Based on samples of the plurality of estimated subband signals (112) and a plurality of decoded subband signals (112) based on samples of the plurality of semi-quantized prediction error subband signals (111) A summation unit configured to determine the samples of the input signal (113); And
And a synthesis filter bank (102) configured to determine an audio signal (114) decoded from the plurality of decoded subband signals (113) using a plurality of synthesis filters. delete delete delete delete delete delete delete delete delete delete delete delete delete delete delete delete delete delete delete

Images